Crop quality & physiology
Test models of grain protein, photosynthesis and root growth against published data and local measurements.
AI × agriculture
A future worth
growing.
We build AI and plant models for better crops, ingredients and production decisions.
Explore the systemsBuild the next
agricultural industry.
Make what plants produce as important as how much they grow. We’re building toward an integrated agricultural company: crop discovery, specified ingredients and resilient US-grown supply.
Today, we build research systems and test them against published data. Physical studies and production come next.
Our research systems
Uno
Decide what is
worth investigating.
How Uno works
Output
A research question and an evidence request
- Input
- A need, a specification, a constraint
- Process
- Compare alternatives, costs, uncertainty and trade-offs
Feedback
Use results to reassess research priorities
Today
Working internally
P14N7
Build a model
that grows with evidence.
Inside P14N7
Foundation
Typed plant records, priors and measurement updates
- Input
- Species knowledge, biological context and local measurements
- Components
- Physiology and growth predictors, adopted engines and prototypes
Next test
Connect carbon, water and organ development to useful compound amount
Today
Local models built; whole-plant integration in development
Century
Make the prediction
answer to evidence.
How Century works
Output
A reusable result, with assumptions and comparison history
- Input
- Observations, model predictions and competing explanations
- Process
- Fit models and test against strong alternatives
Feedback
Keep what works; use failures to choose the next test
Today
Operating research infrastructureInside Century
Compare models of photosynthesis, water loss, temperature, composition and storage against measurements and alternatives.
View model results- Physiology
- Carbon, water & heatCoupled plant and chamber conditions
- Chemistry
- Composition, charge & massExplicit units and balance constraints
- Calibration
- Species knowledge + local assaysLocal calibration has an explicit scope
Keep each measurement connected to its article, table, units and material context. Distinguish observations, formulas and missing cells.
Inspect evidence coverage- Provenance
- Keep the source attachedArticle, table and cell references
- Comparability
- Preserve units & material basisFresh, dry, extracted and recovered differ
- Missing context
- Keep uncertainty explicitExtraction still requires review
Compare competing explanations and design measurements to distinguish them.
See the CPU benchmark- 01Define the outcome that matters
- 02Compare plausible explanations
- 03Find a distinguishing observation
- 04Test it against an ordinary approach
Replay forecasts without refitting. Retain assumptions, comparison history and failed transfers for the next application.
Read the next milestones- Before scoring
- Record the forecast and comparatorTrack what was already known
- After scoring
- Retain the result and its scopeWins, losses and unresolved tests
- Next application
- Reuse the supported componentTest the new context separately
How a result changes the next question
- 01 / Retain
A locally calibrated sunflower predictor held up on a withheld light condition and replayed without refitting.
- 02 / Challenge
A water-and-temperature extension lost to an ordinary method and remained a diagnostic tool.
- 03 / Redirect
Retain the supported predictor; prioritise physiology and growth-to-compound connections.
Published evidence at scale
Atlas · 31 October 2026- Plant measurements read
- 1.81M
- Species represented in source rows
- 8,680
- Open-access articles represented
- 34,726
- Articles screened
- 750,844
Plant measurements read
Compound, trait, element or yield values for one treatment group, extracted from tables in 34,726 open-access articles. Atlas release: 31 October 2026. In a 60-article AI reference check, the parser recovered about half the measurements; about four in five reported values matched.
Machine-extracted measurements from published articles.
About the atlas
- Plant measurements read
- Compound, trait, element or yield values for one treatment group, extracted from tables in 34,726 open-access articles. Atlas release: 31 October 2026. In a 60-article AI reference check, the parser recovered about half the measurements; about four in five reported values matched.
- Species represented in source rows
- Distinct species named on those rows. Rice, wheat, maize, tomato and grape carry the most measurements.
- Open-access articles represented
- Distinct articles represented in the cleaned atlas release.
- Articles screened
- Open-access articles screened; not all yielded a measurement.
Research directions
Plant-derived ingredients
Model madecassoside in Centella for cosmetic ingredients. New-material tests are next, separating plant-line, site, season and laboratory effects.
Algae for cosmetics
Model astaxanthin, algal polysaccharides and fucoxanthin–lipid output across cultivation, recovery and storage.
Colour & storage
Compare models of pigment fading and postharvest chemistry. Turn published-data benchmarks into physical stability studies.
Explore research questions
Question to investigate
Which properties make a biological ingredient useful in a product?
- Define the product requirement
- Compare material properties
- Test application performance
An ingredient comparison tied to a specific application.
Preparing the overview…
Measured progress
Swiss wheat / retrospective comparison
17%lower protein-prediction error
One local assay plus variety history improved forecasts across 43 site-years against the best tested alternative.
RMSE in protein percentage points. Zero-based scale; lower is better.
Wheat comparison & method
| Method | RMSE |
|---|---|
| One assay + variety history | 1.076 |
| Best tested rival | 1.301 |
Error fell 17.3%, rounded to 17%. A retrospective second read within the same AI lineage, with prior exposure to the headline result. Separate organic-wheat and barley comparisons lost to ordinary methods.
Sunflower / physiology
Photosynthesis beyond
the fitted light level.
A locally calibrated model predicted photosynthesis at a withheld light condition and replayed without refitting.
Comparison & method
One plant, one withheld light condition. The predictor beat the training-selected interpolation rival; two unselected smooth alternatives scored better. A water-and-temperature extension did not beat its equal-information control.
Ginseng / growth
Root and whole-plant
mass forecasts.
In one facility’s published data, a model improved root and whole-plant dry-mass forecasts over carry-forward and linear extrapolation.
Comparison & method
Source-exposed temporal development. Shoots effectively tied the carry-forward control; the harvest rule did not beat fixed waiting.
Leaf sensing / measurement
Trait estimation
from spectral measurements.
Spectral models support locally calibrated wheat chlorophyll and water estimates, and nitrogen ranking in a new apple laboratory.
Comparison & method
Retrospective cross-laboratory and fresh-sheet tests using local references. Absolute nitrogen and leaf-mass estimates failed in some new contexts; simpler calibration matched or won some tasks.
+Natural-colour predictionWhere mechanistic, statistical and language models meet
Compare the model
with the alternatives.
Our pigment-stability simulator reduced error against class-average, nearest-analogue and language-model comparisons in a published-data benchmark.
The stronger kinetics and separately sealed comparator remained statistically unresolved, so the all-comparators success criterion was not met.
- Class average
- Lower error
- Nearest analogue
- Lower error
- Language model
- Lower error
- Earlier kinetics
- No clear separation
- Sealed comparator
- No clear separation
+Food-composition screeningA retrospective ranking test
FDA survey / retrospective comparison
0.87cadmium ranking AUC
Ingredient names, label order and food category improved ranking in a later FDA survey over a first-ingredient-plus-category baseline. A screening comparison, not a food-safety test.
386 foods FDA FY2023–24
Training stops in 2022. The advantage does not extend to lead beyond the first ingredient.
Exact values & comparison
| Model | AUC | 95% interval |
|---|---|---|
| Recipe model | 0.87 | 0.81–0.93 |
| Baseline | 0.80 | Not reported here |
Computational efficiency
CPU / experiment selection
83%less CPU for measurement selection
Same recommendations at all 120 checkpoints in a wheat benchmark, with the model and measurement budget fixed.
Reproduced in a separate process.
120 / 120
Paired checkpoints with the same recommendations
20 paired runs. One mark per checkpoint; measurement order sometimes differed.
CPU benchmark & method
October 2026 benchmark: acquisition CPU fell 82.8%, rounded to 83%, across 20 paired runs. Model and measurement budget stayed fixed. Recommendations matched at all 120 checkpoints; measurement order sometimes differed.
Finding the window
worth testing.
How long should a finishing-light treatment last?
A published red-lettuce study lets us test what different models can recover.
Leaf pigment concentrationmg CYA / g dry tissue
The four-day treatment
Both models fit the other eight settings.
- Predicted
- 12.42
- Published measurement
- 12.91
- Linear interpolation
- 8.94
36%
lower prediction error
Quadratic model vs linear interpolation.
Across seven held-out settings in this study.
Cammarisano et al., 2020 ↗Retrospective comparison within one study.
About this comparison
How concentration changes with treatment duration
Leaf pigment concentrationmg CYA / g dry tissue
- Published means
- Quadratic model
- Linear interpolation
Retrospective comparison of a previously inspected red-lettuce study. Each assessed setting was omitted from fitting; the chart shows four days, estimated from the other eight settings.
Across seven interior settings, RMSE fell 36%: 1.026 for the quadratic versus 1.608 for linear interpolation, in mg CYA/g dry tissue. Five models were evaluated; boundary settings were not scored.
CYA denotes cyanidin equivalents, a pigment assay. Points show published means and whiskers show reported standard errors (SEM), not prediction intervals. Four days corresponds to 72 lit hours under an 18-hour daily photoperiod. The duration axis uses log spacing. Treatment timing and cumulative light changed together, so duration was not isolated.
The unmeasured peak is a candidate for testing. The advantage did not extend to fresh mass. Pigment and mass came from separate plants.
What grows.
What it contains.
Five nutrient recipes.
One published lettuce experiment.
Relative to the 0% recipe
- Shoot mass
- Leaf nitrate
Ammonium share of supplied nitrogen
75% ammonium recipe
Each outcome as a share of its 0% recipe value
- Shoot mass
- 96%
- Leaf nitrate concentration
- 36%
Much lower mean nitrate, with mean shoot mass close to the reference.
About this study
Lettuce grown under 24-hour daily light and harvested after four weeks. Dots show approximate published means, each normalized to its own 0% recipe mean. Links connect separate outcomes for one recipe, not paired plant measurements or a combined score.
Approximate values from one reading of Figures 2A and 3B. The paper reports n = 5–6; exact treatment counts are unresolved in this reconstruction.
Other ions and electrical conductivity also varied, so this test does not isolate ammonium. The table gives approximate means ± reported standard errors; uncertainty is not plotted on the normalized chart.
| Ammonium share | Shoot mass (g/plant) | Leaf nitrate (mg/100 g FW) |
|---|---|---|
| 0% | 157.8 ± 8.8 | 373.3 ± 16.0 |
| 25% | 158.7 ± 10.0 | 310.2 ± 7.5 |
| 50% | 159.3 ± 13.4 | 186.1 ± 6.4 |
| 75% | 151.9 ± 8.8 | 135.8 ± 7.0 |
| 100% | 94.3 ± 7.7 | 19.3 ± 3.2 |
From prediction
to production.
Connect growth to useful output.
Connect plant responses, organ growth and compound amount. Test each addition against measurements.
Put forecasts against new measurements.
Run planned studies in fruit chemistry, microalgal pigment and natural-colour stability, with predictions specified in advance.
Develop production around a specification.
Connect harvest, ingredient recovery, stability and resource needs into growing protocols and repeatable supply.
The founding team.
Davis, California
Based in California’s Central Valley - which grows about 40% of U.S. fruit and nuts.
NSF I-Corps
UC Davis AgTech and FoodTech cohort participant
What could we
grow together?
Bring a crop, ingredient or production question. Let’s define the next test.
Email the research team